The last hundred watts a GPU draws buy almost no performance. NVIDIA ships cards tuned to win benchmark charts, not to earn margin at 4 cents a minute. For a hosting rig that pays its own electricity bill, dialing the power limit down is the highest-return five minutes of configuration you will do.
The 80% rule: most of the performance, most of the savings
GPU performance does not scale linearly with power. The voltage and frequency curve flattens hard at the top, so the final 20% of board power typically buys around 5% of throughput. Put the other way: most GPUs keep roughly 95% of their performance at about 80% of stock power. Renters running training or inference jobs rarely notice the difference. Your electricity meter does.
Worked example, one RTX 4090 at $0.15 per kWh:
| Setting | Draw under load | Cost per month (24/7) |
|---|---|---|
| Stock 450 W | 450 W | $48.60 |
| Limited to 360 W (80%) | 360 W | $38.88 |
| Saved | 90 W | $9.72 per card |
On a four-card rig that is close to $39 a month back in your pocket for a few percent of compute, plus less heat for your fan curve to fight and less stress on power delivery. If your break-even price was built on stock wattage, the limit also lowers the floor you can afford, which is worth rechecking in your price floor math.
Set a power limit with nvidia-smi
Two commands. Enable persistence mode so the driver holds settings while no process is using the card, then set the limit in watts:
sudo nvidia-smi -pm 1
sudo nvidia-smi -pl 360Check what the card accepts first: nvidia-smi -q -d POWER shows the current, default, minimum, and maximum limits. Each model has a hard minimum; a 4090 will not go below 150 W no matter what you ask for. Verify the new limit took effect with a plain nvidia-smi, which shows it in the power column.
On a mixed rig, apply different limits per card with the index flag: sudo nvidia-smi -i 0 -pl 360 targets only GPU 0. Without -i, the limit applies to every card in the machine, which is what you want when the rig is all one model.
Start at 80%, then measure
Power limits reset on reboot: persist them
This is the part most hosts miss. nvidia-smi -pl is not saved anywhere. The first unplanned reboot, and every card silently goes back to stock draw. Your margin math is wrong again and nothing tells you.
Two solid ways to make the limit stick:
- Cron watchdog. A small script queries the current limit, compares it to your target, and reapplies it if they differ. Run it from cron once a minute and the limit survives reboots, driver restarts, and anything else. The complete script and crontab line are in the power limit cron job guide in our open source Linux tools collection.
- Systemd unit. A one-shot service that enables persistence mode and applies the limit at boot, after the driver is loaded. Cleaner ordering, one execution instead of a recurring check. The setup is in the power limit guide.
Either works. The cron watchdog is harder to kill; the systemd unit is tidier. Pick one and confirm after your next reboot that nvidia-smi still shows your number.
Does a power limit hurt your rentals?
The honest answer: slightly, and renters almost never notice. A few points to keep it that way:
- Marketplace benchmarks reflect the machine as tested. If you limit power after benchmarking, heavy compute workloads may land a few percent under the advertised numbers. Keep the cut modest (80% or above) and this stays within noise.
- Memory-bound workloads, which include a lot of inference, barely react to a core power limit at all.
- Lower power means lower temperatures and steadier clocks, which can actually improve sustained throughput in a dense, hot chassis.
Next: know your true margin
A power limit only pays off if you know what your hours cost. Feed your electricity rate and hardware into the Power Cost Calculator to see the hourly, daily, and monthly numbers behind every rental, then set your floor from data instead of vibes.